arazzo: 1.0.1
info:
  title: Microsoft Windows 10 Windows ML Load and Evaluate Model
  summary: Load an ONNX model, create a hardware-accelerated evaluation session, and run an inference against bound inputs.
  description: >-
    A local inference flow built on the Windows.AI.MachineLearning namespace. The
    workflow loads an ONNX model from a file path, creates a LearningModelSession
    bound to a chosen device (CPU, DirectX, or DirectXHighPerformance), and
    evaluates the model with a map of input feature bindings, returning the
    output tensors. Every step inlines its request and documents the status code
    it keys on so the flow can be executed without opening the OpenAPI
    description.
  version: 1.0.0
sourceDescriptions:
- name: mlApi
  url: ../openapi/microsoft-windows-10-ml-openapi.yml
  type: openapi
workflows:
- workflowId: load-and-evaluate-model
  summary: Load an ONNX model, create a session, and evaluate it against input bindings.
  description: >-
    Loads an ONNX model from a file path, creates an evaluation session on the
    requested device, and evaluates the model with the supplied feature
    bindings.
  inputs:
    type: object
    required:
    - filePath
    - bindings
    properties:
      filePath:
        type: string
        description: Path to the ONNX model file to load.
      deviceKind:
        type: string
        description: Device to use for evaluation.
        enum:
        - Default
        - Cpu
        - DirectX
        - DirectXHighPerformance
        - DirectXMinPower
        default: Default
      bindings:
        type: object
        description: Map of input feature names to their input values.
      correlationId:
        type: string
        description: Correlation id for tracking the evaluation.
  steps:
  - stepId: loadModel
    description: >-
      Load the ONNX model from the supplied file path and capture the assigned
      model id and feature counts.
    operationId: loadMLModel
    requestBody:
      contentType: application/json
      payload:
        filePath: $inputs.filePath
        modelFormat: ONNX
    successCriteria:
    - condition: $statusCode == 201
    outputs:
      modelId: $response.body#/id
      modelName: $response.body#/name
      inputFeatureCount: $response.body#/inputFeatureCount
  - stepId: createSession
    description: >-
      Create an evaluation session for the loaded model bound to the requested
      device kind.
    operationId: createMLSession
    requestBody:
      contentType: application/json
      payload:
        modelId: $steps.loadModel.outputs.modelId
        deviceKind: $inputs.deviceKind
    successCriteria:
    - condition: $statusCode == 201
    outputs:
      sessionId: $response.body#/id
  - stepId: evaluate
    description: >-
      Evaluate the model with the supplied input bindings and capture the output
      tensors and whether the evaluation succeeded.
    operationId: evaluateMLModel
    parameters:
    - name: sessionId
      in: path
      value: $steps.createSession.outputs.sessionId
    requestBody:
      contentType: application/json
      payload:
        correlationId: $inputs.correlationId
        bindings: $inputs.bindings
    successCriteria:
    - condition: $statusCode == 200
    - context: $response.body
      condition: $.succeeded == true
      type: jsonpath
    outputs:
      succeeded: $response.body#/succeeded
      outputs: $response.body#/outputs
      correlationId: $response.body#/correlationId
  outputs:
    modelId: $steps.loadModel.outputs.modelId
    sessionId: $steps.createSession.outputs.sessionId
    succeeded: $steps.evaluate.outputs.succeeded
    evaluationOutputs: $steps.evaluate.outputs.outputs